🔢 mathematics

Almost sure bounds for weighted sums of Rademacher random multiplicative functions

This paper establishes almost sure upper and lower bounds for weighted sums of Rademacher random multiplicative functions, demonstrating that their growth is dominated by multiplicative chaos and differs significantly from the Steinhaus case, while also providing a sharper bound for sums restricted to integers with large prime factors.

Original authors: Christopher Atherfold

Published 2026-02-04

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Christopher Atherfold

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A Game of Chance with Numbers

Imagine you are playing a game with a giant bag of numbers. For every prime number (like 2, 3, 5, 7...), you flip a fair coin.

Heads: You assign the number +1.
Tails: You assign the number -1.

Once you've assigned these values to the primes, you create a rule for all other numbers: if a number is made of distinct prime factors (like $6 = 2 \times 3$ ), you multiply their assigned values together. If a number has a repeated prime factor (like $4 = 2 \times 2$ ), its value is 0.

This creates a random sequence of numbers, $f(n)$ , that looks chaotic but follows strict mathematical rules. Mathematicians have long been fascinated by the "partial sums" of these numbers—basically, adding them up as you go ( $f(1) + f(2) + f(3) + \dots$ ).

The Problem: The "Weighted" Sum

In this paper, the author isn't just adding the numbers up; he is adding them up with a weight. Specifically, he is looking at the sum:
$\frac{f(1)}{\sqrt{1}} + \frac{f(2)}{\sqrt{2}} + \frac{f(3)}{\sqrt{3}} + \dots + \frac{f(x)}{\sqrt{x}}$

Think of this like a race where the runners (the numbers) get slower and slower as the race goes on. The first runner runs at full speed, the second at half speed, the third at roughly 57% speed, and so on. The question is: How far can this total distance deviate from zero?

In the world of pure randomness, you might expect the sum to stay relatively small, bouncing around zero like a drunk person walking home. However, because these numbers are "multiplicative" (connected by their prime factors), they don't behave like a simple coin toss. They have hidden patterns that can cause massive, unexpected swings.

The Main Discovery: A Surprising Size Limit

The author proves a specific limit on how big these swings can get.

The Result:
For almost every possible outcome of the coin flips, the total sum will never grow larger than a very specific, slow-growing function: roughly $(\log \log x)^{3/4}$ .

The Analogy:
Imagine you are measuring the height of a wave in the ocean. You might expect the waves to be huge. But this paper says, "If you look at almost all possible ocean states, the waves will never exceed a height of about 10 meters, even as the ocean gets infinitely large."

The author also proves a "lower bound": there are definitely times when the sum gets as large as $(\log \log x)^{-1/2}$ . This confirms that the sum doesn't just sit still; it definitely moves, but it has a "ceiling" on how high it can jump.

Why This Is Different from Other Models

Mathematicians often use two types of random models to study these number patterns:

The "Steinhaus" Model: Imagine the coin flips can land on any point on a circle (complex numbers). This is like a spinner that can point anywhere.
The "Rademacher" Model (This Paper): The coin flips are strictly +1 or -1. This is a simpler, more rigid model.

The Twist:
In the "Steinhaus" (spinner) model, the biggest jumps in the sum come from a specific type of mathematical structure called an "Euler product" (a giant multiplication of terms). It's like the biggest waves are caused by a specific wind pattern.

However, in the "Rademacher" (coin flip) model, the author found that the biggest jumps come from a different source entirely: Multiplicative Chaos.

Analogy: In the spinner model, the waves are caused by a steady, predictable wind. In the coin flip model, the waves are caused by a chaotic, turbulent storm where the randomness itself creates the biggest spikes. The "deterministic" part (the predictable math) actually cancels things out more effectively here, leaving the chaotic randomness to dominate the size of the sum.

The "Large Prime" Shortcut

The paper also looks at a specific subset of numbers: those that have a very large prime factor (larger than the square root of the total count).

The Finding:
When you restrict the game to only these "large prime" numbers, the sum behaves much better. The limit drops significantly to roughly $(\log \log x)^{1/4}$ .

Analogy:
Imagine you are trying to predict the weather. If you look at all weather patterns, it's chaotic. But if you only look at days where a massive hurricane is present, the patterns become much more predictable and smaller in scale. The "large prime" numbers act like a filter that removes the most chaotic noise, revealing a cleaner, smaller limit.

Why This Matters (According to the Paper)

The author connects this to the famous Riemann Hypothesis, one of the biggest unsolved problems in math. The Riemann Hypothesis is related to how the "Möbius function" (a cousin of our random $f(n)$ ) behaves.

The Reality Check: The author suggests that while our random coin-flip model ( $f(n)$ ) is a good way to study these sums, it might not perfectly predict the behavior of the real Möbius function. The random model shows that the sums can get quite large (due to the "chaos"), which suggests that if the real Möbius function behaves similarly, it might be harder to prove the Riemann Hypothesis than we thought.

Summary of the "Tools" Used

To prove these limits, the author didn't just guess. He used a "toolbox" of advanced probability techniques:

Martingales: A way of tracking a random walk where the future depends only on the present, not the past. He treated the sum as a series of steps where he could predict the "variance" (how much it might wiggle).
Splitting the Problem: He broke the massive sum into smaller chunks based on the size of the prime factors (small primes vs. large primes) and analyzed them separately.
Euler Products: He looked at the sum as a giant multiplication problem and used complex calculus (integrals) to measure the "energy" of the system.

The Bottom Line

Christopher Atherfold has drawn a precise map of the "wilderness" of these random number sums. He proved that while the sums can wander far, they are strictly bounded by a specific, slow-growing limit. He also discovered that the "wildness" comes from a different source in this coin-flip model compared to other mathematical models, highlighting that the simple act of flipping a coin (+1/-1) creates a unique type of mathematical chaos.

Technical Summary: Almost Sure Bounds for Weighted Sums of Rademacher Multiplicative Functions

Problem Statement
The paper investigates the almost sure behavior of the weighted partial sums of Rademacher random multiplicative functions, defined as:
$M_f(x) := \sum_{n \le x} \frac{f(n)}{\sqrt{n}}$
where $f(n)$ is a Rademacher multiplicative function (independent Rademacher variables on primes, extended multiplicatively to square-free integers, and zero otherwise). This problem serves as a probabilistic model for understanding the cancellation properties of the Möbius function $\mu(n)$ , specifically regarding the Riemann Hypothesis.

The study contrasts with the "Steinhaus" case (where $f(p)$ are uniform on the unit circle), where previous work by Hardy established bounds resembling the Law of the Iterated Logarithm (LIL). The central question is whether the Rademacher case exhibits similar fluctuation sizes or if the deterministic structure of the Rademacher Euler product leads to different asymptotic behavior.

Methodology
The proof strategy synthesizes techniques from Caich (unweighted Rademacher bounds), Hardy (weighted Steinhaus bounds), and Mastrostefano (large prime factor restrictions). The methodology proceeds through several key stages:

Reduction to Test Points: Following Lau–Tenenbaum–Wu and Caich, the continuous variable $x$ is reduced to a sparse sequence of test points $x_i = [e^{i\gamma}]$ . This allows the use of the first Borel–Cantelli lemma, provided the increments between test points are controlled.
Prime Factor Splitting: The sum $M_f(x_i)$ $M_{f} (x_{i})$ is decomposed based on the size of the largest prime factor $P(n)$ $P (n)$ of the integers $n$ $n$ . A sequence of cutoffs $y_j$ $y_{j}$ is introduced to split the sum into:
- Smooth numbers ( $P(n) \le y_0$ ).
- Sums over ranges $y_{j-1} < P(n) \le y_j$ .
  This decomposition reveals a martingale difference structure, allowing the application of Doob's inequalities and Azuma–Hoeffding bounds.
Euler Product Analysis: The core of the upper bound relies on analyzing the Rademacher Euler product:
$F_y(s) = \prod_{p \le y} \left(1 + \frac{f(p)}{p^s}\right)$
Unlike the Steinhaus case, the Rademacher product lacks translation invariance. The paper utilizes Harper's "Multiplicative Chaos" results and Gaussian random walk approximations to bound the moments of these products.
Handling Deterministic vs. Random Contributions: A critical distinction in the Rademacher case is the presence of a deterministic negative term in the logarithm of the Euler product (arising from the expansion of $\log(1+f(p)/p^s)$ ). The paper shows that for $s=1/2$ , this deterministic term dominates the random fluctuations, unlike in the Steinhaus case where the random chaos is the primary driver.
Lower Bound Construction: The lower bound adapts Hardy's method for Steinhaus functions but accounts for the specific structure of the Rademacher product. It utilizes Jensen's inequality on an integral of the Euler product to isolate a deterministic contribution of size $\log \log x$ .

Key Contributions and Results

Theorem 1 (Upper Bound): For any $\epsilon > 0$ , almost surely:
$M_f(x) \ll (\log \log x)^{3/4 + \epsilon}$
This bound is significantly larger than the $(\log \log x)^{1/4}$ prediction one might expect from a direct analogy with the Steinhaus case or the Law of the Iterated Logarithm. The author attributes this to the "multiplicative chaos" contribution becoming dominant due to the specific behavior of the Rademacher Euler product.
Theorem 2 (Restricted Upper Bound): When restricting the sum to integers with a single large prime factor ( $P(n) > \sqrt{x}$ ), the bound improves to:
$\sum_{\substack{n \le x \\ P(n) > \sqrt{x}}} \frac{f(n)}{\sqrt{n}} \ll (\log \log x)^{1/4 + \epsilon}$
This suggests that the larger bound in Theorem 1 arises from the interaction of integers with multiple prime factors (specifically, the difficulty of separating smooth contributions in the weighted case).
Theorem 3 (Lower Bound): There exist arbitrarily large $x$ such that:
$|M_f(x)| \gg (\log \log x)^{-1/2}$
The proof identifies a deterministic contribution in the lower bound analysis that is absent in the Steinhaus setting.
Corollary 1 (Completely Multiplicative Case): The results are applied to the completely multiplicative Rademacher function $f^*(n)$ , yielding:
$\sum_{n \le x} \frac{f^*(n)}{\sqrt{n}} \ll \log x (\log \log x)^{3/4 + \epsilon}$
almost surely.

Significance and Claims
The paper claims to be the first result demonstrating that the Rademacher model provides more cancellation (in the sense of smaller moments for certain orders, though here the upper bound is larger due to the specific weighting) than the Steinhaus model in the context of weighted sums.

The primary significance lies in the identification of the multiplicative chaos contribution as the dominant factor for large fluctuations in the Rademacher case, contrasting with the Steinhaus case where the size of the random Euler product itself drives the fluctuations. The author notes that the upper bound in Theorem 1 is likely not sharp, conjecturing that the true order of magnitude is $(\log \log x)^{1/4 + \epsilon}$ , matching the restricted case and the Steinhaus case. The gap between the proven upper bound and the conjectured sharp bound is attributed to the difficulty of controlling the "smooth" part of the sum (integers without large prime factors) in the weighted setting.

The work relies heavily on the "black box" probability results from Harper's work on low moments and multiplicative chaos, adapting them to the specific constraints of Rademacher variables and weighted sums. The paper does not propose new experimental applications but offers a refined theoretical understanding of random multiplicative functions, potentially informing future quantitative central limit theorems for such sums.